Monday , February 19 2018

Intel Kills Knights Hill, Will Launch Xeon Phi Architecture for Exascale Computing


In 2006, Intel decided to build a new GPU architecture, codenamed Larrabee. It didn’t work, so Intel decided to re-spin it as a manycore architecture dubbed Knights Ferry. While the first KNF processors were prototypes with literal GPU silicon still bolted on-die, follow-up chips like Knights Corner switched to a modified Pentium (P54C) architecture with x86 ISA support, 4-way SMT, and up to 1.2TFLOPS of double-precision performance. The Knights Landing architecture updated the CPU core to a heavily modified version of Intel’s Silvermont uarch, increased the core count up to 72 cores, added MCDRAM (aka Hybrid Memory Cube), AVX-512 support, and was supposed to set the stage for a 10nm refresh with 2nd-generation Omnipath — except Intel just quietly killed it.

You could be forgiven for not noticing, even if you read Intel’s blog post on SC17, the annual supercomputing convention. Out of the company’s 828-word blog post, just 30 of them deal with the cancellation of Knights Hill, and it’s buried in a section labeled “Enabling Exascale.” (PR pro tip: If you want to hide the negative news, put it under a positive header). It reads: “One step we’re taking is to replace one of the future Intel Xeon Phi processors (code name Knights Hill) with a new platform and new microarchitecture specifically designed for exascale.”

Intel also had a part named Knights Mill in development, but we don’t know much about that chip yet, beyond that it supports half-precision work like GPUs from AMD and Nvidia and is intended for AI and deep learning workloads. While we don’t know much about the low-level architecture of Knights Hill, we do know something about what kind of deployment Intel expected the chip to see, via the Aurora supercomputer project.


There are a few different ways to read this announcement, and we don’t know which is accurate yet.

The most direct is that Intel is canceling Knights Hill as a separate product because it believes it can target both spaces (HPC, AI/DL) with the same part. Years ago, we talked about experimental research Intel had done into variable floating point computing, which saved on power by giving programmers the option to choose how much precision to retain on a case-by-case basis.

Second, it’s possible that something has gone badly wrong with some aspect of the chip’s design, to such an enormous degree that cancelling the project and starting over is the only way to fix it. This wouldn’t automatically point to something wrong with the CPU core, which sounded like a reasonable die shrink with some new features — it could be a scaling or power consumption issue with some other aspect of the design. There’s no evidence to suggest this is true, it’s just a technical possibility.

Third, there’s the chance that Intel has decided Knights Hill isn’t going to hit the aggressive performance or capability targets it wants and is re-spinning the design to do so. Xeon Phi products have always had a much longer cadence than other parts in the consumer or even enterprise — Intel has refreshed its Xeon processors much more quickly, for example — and so, it may be thinking long-term about the future of exascale and choosing to take a hit now rather than risk being dethroned later on. Either way, don’t be surprised if Nvidia and even AMD play up this strategic gap with new launches of their own. Hat tip to Tech Report for sussing out Intel’s blink-and-you’ll-miss-it announcement.

About Skype

Check Also

Windows 10 Is Adding an ‘Ultimate Performance’ Mode

Over the last year, Microsoft has been busy adding new SKUs to Windows 10. Windows …

Leave a Reply

Your email address will not be published. Required fields are marked *